OpenClaw stack runs 100% locally

A local artificial intelligence stack means the model, chat history, and tools run on your own machine or server instead of a vendor’s cloud. OpenClaw’s current docs show that setup is now a first-class way to run the assistant. (docs.openclaw.ai) (github.com) OpenClaw describes itself as a personal AI assistant that runs on your own devices and connects to messaging platforms including WhatsApp, Telegram, Slack, Discord, Signal, iMessage, and Microsoft Teams. Its GitHub repository showed about 365,000 stars on April 27, 2026. (github.com) The core piece is the “gateway,” which OpenClaw says owns sessions, authentication profiles, channels, and state. In the default remote design, that gateway binds to loopback and listens on port 18789. (docs.openclaw.ai) Loopback is the computer talking to itself, like sending mail to your own apartment mailbox instead of the street. OpenClaw’s docs say remote users can reach that local-only gateway by forwarding port 18789 over Secure Shell, or SSH. (docs.openclaw.ai) On macOS, OpenClaw’s app now includes a “Remote over SSH” mode that executes commands on another host and reuses the same tunnel for health checks, Voice Wake forwarding, and Web Chat. The docs say the app opens an SSH connection with key-based authentication and a local port-forward. (docs.openclaw.ai) That means the control path can stay on a local network, a home server, or a private tailnet instead of passing through a hosted relay. OpenClaw’s security notes recommend loopback binds on the remote host and connecting through SSH or Tailscale. (docs.openclaw.ai 1) (docs.openclaw.ai 2) The model side works the same way: you download the model weights, which are the files containing what the model has learned, and run them on your own hardware. XDA wrote on April 26 that self-hosted setups using Ollama, LM Studio, and llama.cpp let users swap models depending on the workload. (xda-developers.com) XDA’s examples were specific. It said Qwen 2.5 Coder and GPT-OSS were strong for programming, Llama 3.1 worked well for long notes and research, and Qwen 3.5 plus DeepSeek were better for calling tools on apps like Nextcloud, a network-attached storage server, and Home Assistant. (xda-developers.com) That mix-and-match approach lines up with OpenClaw’s design, where one gateway can broker chats, tools, and remote nodes while the operator chooses where the models run. The result is less about one “best” model and more about keeping state, credentials, and traffic under the operator’s control. (docs.openclaw.ai) (xda-developers.com) The local-first pitch is no longer just hobbyist theory. OpenClaw’s current docs and the latest XDA reporting both describe a setup where the assistant, the tunnel, and the model can all stay close to home. (docs.openclaw.ai) (xda-developers.com)